LLM-Powered Recommendation System

Activity and event recommendation is cold-start heavy and context-rich: catalogs churn, user intent is expressed in natural language, and pure collaborative signals are sparse. The system needs to reason over semantics and preferences, not just co-occurrence.

AI engineer. I designed and built the end-to-end recommendation pipeline, from preference encoding through retrieval, LLM enrichment, and ranking.

1. 01User Profile — structured preferences plus interaction history.
2. 02Preference Encoder — turns profile and context into a query representation.
3. 03Candidate Retrieval — semantic vector search over the activity/event catalog.
4. 04LLM Enrichment — language model reasons over candidates, explanations, and fit against stated intent.
5. 05Ranking — final ordering combining relevance, diversity, and preference signals.
6. 06Results — personalized, explainable recommendations.

• Retrieve-then-rank over end-to-end LLM scoring

  Scoring the whole catalog with an LLM is too slow and expensive. Cheap semantic retrieval narrows to a candidate set; the LLM only reasons over a shortlist.

• Embeddings for recall, LLM for precision

  Vector search maximizes recall over a churning catalog; the LLM adds precision and natural-language justification where it changes the decision.

• Explainability as a first-class output

  For discretionary activities, a short 'why this' rationale materially improves trust and click-through, so the enrichment step emits explanations, not just scores.

Offline evaluation on held-out interactions using retrieval recall@k and ranking quality (NDCG-style), plus qualitative review of LLM rationales for faithfulness. Latency and cost per request are tracked because they gate what LLM stages are affordable in production.

• LLM enrichment can hallucinate attributes not present in catalog metadata; constraining it to provided fields and validating outputs mitigates this.
• Semantic retrieval can over-cluster popular categories, hurting diversity; explicit diversification in ranking counteracts it.
• End-to-end latency is dominated by the LLM stage, forcing careful shortlist sizing and caching.

• LLMs
• Embeddings
• Vector Search
• Ranking
• Python